Thief Week - AI Feature

With the eagerly awaited release of Thief: Deadly Shadows only days away, TVG has teamed up with Eidos and Ion Storm to bring you a whole â??taffingâ? load of information on the latest and last title in the acclaimed stealth genre.

Today itâ??s the turn of David Riegel Programmer Ion Storm, to discuss the complexities of the AI featured within the game and the balance between stealth and action.

Thief: Deadly Shadows â?“ Developer Feature AI

There were a lot of really interesting decisions that went into planning how the AI in Thief: Deadly Shadows were going to respond and interact with their environment. We knew early on that we were going to be writing the AI code from scratch, so we had an opportunity to think about the core stealth experience in Thief 1 and Thief 2, what was wrong with it, and how we could improve upon it. Thief: DS required a lot of unique features, not present in Thief 1 and 2, in order to fulfill the vision of the game â?“ non-hostile AIs, a faster and more advanced combat system, factional relationships, etc. However, our ultimate focus rested on the core stealth experience. The biggest questions on our minds were, firstly, how we could maintain stealth tension whilst retaining an action feel and, secondly, how we could improve AIsâ?? spatial awareness of their surroundings during searching.

So, right away it was down to business: we spent a bunch of time looking at stealth games on the market and of course analyzing Thief 1 and 2. Actually, we spent a lot of time playing other stealth gamesâ?¦ definitely the hard part of the research!

Interestingly, it was at first sometimes difficult to determine what made some stealth games work and what made some fail. We did notice trends throughout the entire genre though. One of the things we noticed right away is that in stealth-action games the stealth and the action are often two independent features. The stealth element is the slow paced, sneaking, tension gameplay, and the action is what happens when you get caught. The games that tended to be the most successful were the ones who were good at creating both good combat and good stealth. We also noticed that the only game to have a stealthy nekkid dude, MGS 2, was hugely successful.

Ok, so we wanted good combat and good stealth, no problem. But how could we take the core stealth itself and make it better? Too often itâ??s easy to create stealth that is too slow and not as satisfying as it could be. Each individual element of Thief: DS had to be strong in and of itself and not just the combination, much likeâ?¦ well much like twoâ?¦ strongâ?¦ things. One of the elements that is fun about traditional pure stealth is that it creates moments of fantastic suspense â?“ lurking in darkness or hiding behind objects while an AI searches for you is part of the core fantasy element. And whatâ??s not fun you might ask? Well, to us it was the wait for long periods of time part. Those moments of tension created when youâ??re hiding in shadows are pure adrenaline, waiting to see if youâ??re going to get caught â?“ but after a certain amount of time that tension fades and it becomes a waiting contest. You have to wait for the AI to finish what itâ??s doing before you can move again. Suspense is good, but an elevator-riding simulation we are not. One of our design goals was to keep the player tense, but not confined to a single dark space â?“ we wanted things moving and fluid â?“ at least a large part of the time.

At that point we began to consider the problem of spatial awareness. When doing searching behavior in stealth games, itâ??s really easy to create obsessive compulsive AI. No, not the fun kind that wash their hands a million times and turn the oven on and off â?“ those are fine. I mean the kind of AI that walks around the same space, checks the same rooms, and generally doesnâ??t really understand what kind of space itâ??s in. We knew that we wanted AI that were more spatially aware and could tell where theyâ??d searched versus where they hadnâ??t.

After some planning we discovered we could solve both the issue of action based stealth and of spatial awareness at the same time. Being confined to a single space and waiting was often caused by AIs in games who simply would not leave the area they were searching. Now, back when I was a fanboy and I was playing Thief 1, I didnâ??t want to leave many of those cool rooms either â?“ hey, they were cool! â?“ but to kick the AI in Thief: DS up a notch, we needed more than garlic â?“ we were going to have to give them a more robust spatial intelligence. If an AI was going to search an entire room and not just part of it, keeping the player on their toes, it would not only seem more human, but would also give rise to some interesting, action oriented player behavior. We came up with a plan; if an AI was alerted when it was on one side of the room, we wanted it to cautiously move from one side of the room to the other until it was satisfied the space was clear of intruders. If an AI was alerted in the middle of a room, we wanted it to move to one side, search, then move quickly through the middle of the room, then slowly search the other side of the room. If an AI was really suspicious and wasnâ??t satisfied when the whole room was cleared, we wanted it to move on to an adjacent room or poke its head through a door.

The real challenge came in making the AI predictable enough to preserve the gameplay, but human enough to be thorough in all of its searches. To help fulfill this, we decided not to make the searching entirely time based â?“ time is a factor, but an AIâ??s searching behavior is also informed by what it has to search versus what ground itâ??s already covered. At any given time, an AI can be interrupted by new evidence (sight and sound), other alert AIs, or changes in the environment. This means that an AI wonâ??t give up until itâ??s done with an area and knows nothing funny is going on. Not funny haha, but jump up and stab-you-in-the-back funny. This kind of searching behavior can be challenging, but it also means you can distract AIs and set traps for them with way cool player tools - the AI will be smart enough to investigate your distractions and, if youâ??re careful, dumb enough to fall for your traps. Further, it means if an AI is in a small space, it wonâ??t repeat the same behaviors over and over again, but will give up more quickly as you would expect.

One of the things people remember from Thief 1 and 2 are the guard personalities. In Thief: DS, we wanted to keep them grumbling about all of the important stuff â?“ dinner, the sir, how theyâ??re gonna bonk some taffers over the head â?“ but we needed to make sure they were broadcasting intelligently to correspond to their new behaviors too. Having them say â??Hey, I think Iâ??ll go look behind those cratesâ? or â??Hey, who left that door open, Iâ??m going to go check it outâ? allows the player to make some predictions about how theyâ??re searching and what theyâ??re doing before they give up. Thief players who are as deadly as they are shadowed will be able to take advantage of this â?“ theyâ??ll be able to sneak up behind guards and backstab or blackjack them as the guards go to investigate new areas or turn their backs. It works both ways though â?“ if a guard starts coming your way you have to find a new shadow, make some shadow of your own, or lure the guard into a mine or oil puddle. Fun stuff!

Balancing out this new approach whilst keeping enough of the tense hiding moments was pretty tough, but in the end they seemed to come together pretty well. We tried to get as many of the â??Oh crap!â? tense hiding situations as we could and as much potential for the â??Saianora taffer!â? stealthy action moments as humanly possible . Did we manage to accomplish everything we set out to accomplish with regards to stealth? I like to think soâ?¦ except for maybe the nekkid cartwheels, but we can save that for another game.

TVG will be back tomorrow with more information on the development of Thief: Deadly Shadows along with the chance of picking up yourself a free copy.